Qwen3.6-27B hits 20 TPS on budget hardware

The Qwen3.6 architecture proves that dense models can overcome the quadratic attention bottleneck, making 27B-parameter models viable for high-speed local inference.

• –Gated DeltaNet (GDN) linear attention reduces memory usage by 80%, allowing 20+ TPS on workstations with as little as 8GB VRAM and offloaded system RAM.
• –Thinking Preservation retains reasoning traces across multi-turn conversations, drastically reducing re-computation time for complex repository-level coding tasks.
• –Performance matches Claude 4.5 Opus on Terminal-Bench 2.0, providing a top-tier open-source alternative for developers prioritizing privacy and local control.
• –Native support for massive contexts up to 1M tokens paired with linear efficiency allows for comprehensive analysis of large codebases without significant slowdown.
• –Apache 2.0 licensing and immediate GGUF support integrate seamlessly into existing local LLM ecosystems like llama.cpp and KTransformers.